整个函数定义在 chat_agent_executor.py:278，最核心的"搭图"动作集中在函数尾部 chat_agent_executor.py:860-1003。先看它建图的三行主干：

# chat_agent_executor.py:860
workflow = StateGraph(
    state_schema=state_schema or AgentState, context_schema=context_schema
)
# :864 两个核心节点
workflow.add_node("agent", RunnableCallable(call_model, acall_model), input_schema=input_schema)
workflow.add_node("tools", tool_node)
# :884 入口
workflow.set_entry_point(entrypoint)   # 默认 entrypoint = "agent"

图要跑，得先有"状态 schema"。默认状态定义在 chat_agent_executor.py:57：

# chat_agent_executor.py:57
class AgentState(TypedDict):
    """The state of the agent."""
    messages: Annotated[Sequence[BaseMessage], add_messages]
    remaining_steps: NotRequired[RemainingSteps]

# :69  Pydantic 版本（同字段，默认 remaining_steps=25）
class AgentStatePydantic(BaseModel):
    messages: Annotated[Sequence[BaseMessage], add_messages]
    remaining_steps: RemainingSteps = 25

入口签名 chat_agent_executor.py:278，裁剪掉类型噪音看关键参数：

# chat_agent_executor.py:278
def create_react_agent(
    model,                       # 模型：实例 / "openai:gpt-4" 字符串 / 动态可调用
    tools,                       # 工具：list[BaseTool|Callable|dict] 或一个 ToolNode
    *,
    prompt=None,                 # 系统提示：字符串 / SystemMessage / 函数 / Runnable
    response_format=None,        # 结构化输出 schema（额外加一个节点）
    pre_model_hook=None,         # 调模型前的钩子节点
    post_model_hook=None,        # 调模型后的钩子节点
    state_schema=None,           # 自定义状态（默认 AgentState）
    context_schema=None,         # 运行时上下文类型
    checkpointer=None,           # 检查点（记忆，D33）
    store=None,                  # 长期记忆 Store（D40）
    interrupt_before=None,       # 静态断点（D43）
    interrupt_after=None,
    version="v2",                # v1=一次性去 tools；v2=Send 并行分发
    name=None,
    ...
) -> CompiledStateGraph:

建图前，函数先把 tools 参数规整成一个 ToolNode（chat_agent_executor.py:553）：

# chat_agent_executor.py:553
llm_builtin_tools: list[dict] = []
if isinstance(tools, ToolNode):
    tool_classes = list(tools.tools_by_name.values())
    tool_node = tools                                   # 你直接传了 ToolNode，就用它
else:
    llm_builtin_tools = [t for t in tools if isinstance(t, dict)]   # dict = 模型内建工具
    tool_node = ToolNode([t for t in tools if not isinstance(t, dict)])  # 其余包成 ToolNode
    tool_classes = list(tool_node.tools_by_name.values())

而 "agent" 节点里是 call_model（chat_agent_executor.py:661），它做三件事，逐行 D54 讲，这里只看骨架：

# chat_agent_executor.py:661（裁剪）
def call_model(state, runtime, config):
    model_input = _get_model_input_state(state)          # ① 取消息、校验历史
    response = static_model.invoke(model_input, config)  # ② 调模型（含 prompt+bind_tools）
    response.name = name
    if _are_more_steps_needed(state, response):          # ③ 步数不够就返回"抱歉"
        return {"messages": [AIMessage(id=response.id, content="Sorry, need more steps...")]}
    return {"messages": [response]}                      # 正常：把 AIMessage 追加进状态

两个节点靠一条条件边连起来。路由函数 chat_agent_executor.py:831：

# chat_agent_executor.py:831
def should_continue(state) -> str | list[Send]:
    messages = _get_state_value(state, "messages")
    last_message = messages[-1]
    # 模型没要求调工具 → 结束（或去钩子/结构化输出节点）
    if not isinstance(last_message, AIMessage) or not last_message.tool_calls:
        if post_model_hook is not None:
            return "post_model_hook"
        elif response_format is not None:
            return "generate_structured_response"
        else:
            return END
    # 模型要求调工具 → 去 tools
    else:
        if version == "v1":
            return "tools"
        elif version == "v2":
            if post_model_hook is not None:
                return "post_model_hook"
            return [                                  # ★ v2：每个 tool_call 一个 Send
                Send("tools", ToolCallWithContext(
                    __type="tool_call_with_context", tool_call=call, state=state))
                for call in last_message.tool_calls
            ]

骨架有了，函数再把可选件按需接上，最后 compile。看连边与编译（chat_agent_executor.py:872-1003）：

# chat_agent_executor.py:872  pre_model_hook 让入口前移
if pre_model_hook is not None:
    workflow.add_node("pre_model_hook", pre_model_hook)
    workflow.add_edge("pre_model_hook", "agent")
    entrypoint = "pre_model_hook"
else:
    entrypoint = "agent"
# :966  agent 出来走 should_continue
workflow.add_conditional_edges("agent", should_continue, path_map=agent_paths)
# :988  tools 跑完回到入口（无 return_direct 时）
workflow.add_edge("tools", entrypoint)
# :995  编译，透传所有记忆/中断配置
return workflow.compile(
    checkpointer=checkpointer, store=store,
    interrupt_before=interrupt_before, interrupt_after=interrupt_after,
    debug=debug, name=name,
)

✋ 10 分钟动手

# 1. 通读建图段（骨架就这 140 行）
sed -n '553,561p'  libs/prebuilt/langgraph/prebuilt/chat_agent_executor.py
sed -n '831,1003p' libs/prebuilt/langgraph/prebuilt/chat_agent_executor.py

# 2. 亲手看它画出来的图
python - <<'PY'
from langgraph.prebuilt import create_react_agent
from langchain_core.tools import tool
@tool
def add(a:int,b:int)->int:
    "加法"
    return a+b
# 用假模型也行，这里只看图结构
agent = create_react_agent("openai:gpt-4o-mini", [add])
print(agent.get_graph().draw_ascii())   # 打印 START/agent/tools/END 拓扑
PY